Home › Blog › Why photo restoration became a commodity
Why photo restoration became a commodity
Ten years ago, restoring a damaged photo meant paying a skilled retoucher for an afternoon. Today the same fix costs cents of computer time, and dozens of apps sell it. This article explains what happened, in three steps, and what stubbornly refused to get cheap. Spoiler: the AI model is no longer the difference between apps. Judgment and trust are.
2015: you paid for someone's afternoon
Before AI touched this field, restoration was handwork. A retoucher opened your scan in Photoshop and rebuilt the torn corner stroke by stroke, the way a painting conservator works, hours per print, priced like the craft it was. Scanning a shoebox was cheap; fixing what was inside it was not.
2016 to 2021: the robot toolbox era
Then researchers taught neural networks the individual jobs. One network guessed colors for black and white photos. Another rebuilt faces; another sharpened; Microsoft wired several into a pipeline for old photos specifically. The labor moved from retouchers to machine-learning teams, but it was still labor: to ship an app you trained, hosted, and babysat a whole toolbox of models. Restoration got cheaper per photo, but building a restoration product stayed expensive, so there were few of them.
2022: one model learned the whole job
Then diffusion
models arrived: networks that learn what photographs look like by studying
millions of them, and can produce a clean image that fits a damaged one. (Our
companion article explains how
that works.) The important part for the economics: research kept showing that
one strong general model could do what the whole toolbox did:
colorize and fill
holes, repair
without task-specific training,
handle whole families
of damage. With latent
diffusion the compute cost dropped an order of magnitude, and with
instruction
following the toolbox became a sentence: "remove the scratches, keep the
faces."
2023 onward: renting beat building
One barrier remained: running these models takes serious GPUs, and running them reliably takes an infrastructure team. Hosted inference platforms removed that too. Today a state-of-the-art image model is an internet call priced in cents. The capability that needed a research lab in 2020 needs a vendor account in 2026.
That is the moment restoration became a true commodity. When anyone can rent
the same model for cents, the model cannot be what makes an app good, and the
app stores filling up with near-identical restoration apps are the proof.
What never became cheap
Three things did not commoditize, and they are now the whole difference between restoration apps.
Judgment before generation. A good restoration starts with reading the photo: what is damaged, what should be preserved, what the model should be told. Apps that skip this produce the over-smoothed, waxy look you have seen.
Honesty at the limits. These models return a plausible image, not the truth. Where damage destroyed a face, an honest product says so instead of quietly inventing a stranger. That is a policy, not a model property.
Trust with your originals. People upload irreplaceable photos.
Where they are stored, how fast they are deleted, whether they train someone's
next model: no rented API answers that for you. Every app decides, and most
never tell you what they decided.
What this means if you have a box of photos
Trying is now free, so try before you pay anyone: the commodity layer alone
often produces a result that would have cost real money a decade ago. Then
judge apps on the top layer: do they tell you their limits, and do they tell
you what happens to your photos? Ours reads each photo first and writes the
model's instructions automatically, tells you that faces beyond saving are
beyond saving, and deletes your photos after processing. The first restoration
is free in your browser. No download. Photos are deleted from our servers after processing.
The paper trail
- Zhang et al., Colorful Image Colorization, 2016. arXiv:1603.08511
- Wan et al., Bringing Old Photos Back to Life, 2020. arXiv:2004.09484
- Ho et al., Denoising Diffusion Probabilistic Models, 2020. arXiv:2006.11239
- Wang et al., GFP-GAN, 2021. arXiv:2101.04061
- Wang et al., Real-ESRGAN, 2021. arXiv:2107.10833
- Saharia et al., Palette, 2021. arXiv:2111.05826
- Rombach et al., Latent Diffusion Models, 2022. arXiv:2112.10752
- Lugmayr et al., RePaint, 2022. arXiv:2201.09865
- Kawar et al., DDRM, 2022. arXiv:2201.11793
- Brooks et al., InstructPix2Pix, 2023. arXiv:2211.09800
- Zhang et al., ControlNet, 2023. arXiv:2302.05543